Paper submitted to EAGE Annual 2026
Summary
Velocity Model Building (VMB) is traditionally a complex and time-consuming process that requires highly experienced geophysicists skilled in technologies such as Full Waveform Inversion (FWI), various forms of depth imaging, geological interpretation and tomography. While FWI is a powerful tool for constructing accurate velocity models, its success is often heavily dependent on the quality of the initial velocity model. Furthermore, in frontier exploration areas, a robust initial model is generally unavailable, sometimes relying only on legacy 2D data. This reliance on extensive human intervention and the inherent difficulty of acquiring accurate initial models creates bottlenecks in the VMB sequence, which deep learning methods are well-positioned to address. The application of Machine Learning techniques to the task of Velocity Estimation has a long history (e.g. Roth & Tarantola (1994), Calderon-Macias et al (1998) and Araya-Polo et al (2018)). This study seeks to close the gap between aspirational work and pragmatic steps to delivering workflows that generalize on real unseen datasets and focuses on leveraging deep learning techniques to significantly reduce the turn-around time for VMB. We focus on three key areas where considerable time is currently spent: estimating a robust initial macro-velocity model, accurately picking complex water-bottom horizon, and replacing tomography and interpretation steps for automating an accurate FWI initial model.